Browser and desktop tools
Understand when an agent can read public pages and when it needs a connected control tool.
Match the tool to the action
Public web search and page reading can gather information. Signing in, clicking a private dashboard, working with local files or controlling desktop software needs a connected tool that actually supports that action.
Services such as browser MCP servers or Desktop Commander may expose different capabilities depending on their setup. Read the connected server’s tool catalog rather than assume its name guarantees access to your browser or computer.
Test a bounded action
- Connect the appropriate server and inspect its authentication and available tools.
- Give the agent a specific target and a small action, such as reading a test page or listing a permitted folder.
- Review any permission request and the scope of resources shared.
- Check the result in the actual browser, folder or application.
- Expand access only if the next task needs it.
Define operational boundaries
Tell the agent which account, files and environment it should use, and when to ask you. A browser session can contain private data and accounts unrelated to the task; connection alone is not permission to act on all of them.
If a server exposes only read operations, the agent cannot add write access by asking differently. Resolve a missing capability through the server’s documented setup and permissions.